OpenAI introduced GPT-5.5 on April 23 as a model built for agents that plan, use tools, inspect their own work and complete longer tasks with fewer human corrections. It is the company’s first retrained base model since GPT-4.5 and was co-designed with Nvidia GB200 and GB300 NVL72 infrastructure.
The model is rolling out through ChatGPT and Codex for Plus, Pro, Business and Enterprise users, with API access following on April 24. Its benchmark gains are strongest in command-line execution, software work and long-context retrieval.
Higher capability comes at a higher price, making task completion—not cost per token—the relevant enterprise measure.
Agentic benchmarks show substantial gains
GPT-5.5 scores 82.7% on Terminal-Bench 2.0, compared with 75.1% for GPT-5.4 and 69.4% for Claude Opus 4.7. The benchmark tests command-line workflows requiring planning and coordinated tool use inside a sandbox.
On SWE-Bench Pro, it reaches 58.6% for GitHub issue resolution. OpenAI’s internal Expert-SWE benchmark contains tasks with a median estimated human completion time of 20 hours; GPT-5.5 scores 73.1%, up from 68.5% for GPT-5.4.
Long-context retrieval shows an even larger change. On MRCR v2 at one million tokens, GPT-5.5 reaches 74%, versus 36.6% for GPT-5.4.
The comparison is incomplete. Claude Opus 4.7 leads Scale AI’s MCP Atlas tool-use benchmark at 79.1%, while GPT-5.5 has no recorded result. Teams building heavily around MCP should test that gap directly.
Token efficiency softens but does not erase pricing
API pricing is $5 per million input tokens and $30 per million output tokens, twice GPT-5.4’s rates. OpenAI argues that fewer tokens are needed to complete the same Codex task, leaving effective cost about 20% higher. Artificial Analysis reportedly validated that efficiency claim.
GPT-5.5 Pro costs $30 per million input tokens and $180 per million output tokens. It uses additional parallel test-time compute for difficult problems and scores 90.1% on OpenAI’s BrowseComp web-browsing benchmark.
At ten million output tokens monthly, standard GPT-5.5 costs $300 versus $250 for Claude Opus 4.7. The premium pays off only if stronger completion reduces retries, supervision and failed runs.
Enterprises should measure cost per accepted outcome on their own workflows. A model using fewer tokens can still be uneconomic if tools, latency or human review remain expensive.
Production use will determine the real advantage
OpenAI says more than 85% of its employees use Codex weekly across functions including engineering and marketing. Its communications team used GPT-5.5 to analyse six months of speaking requests and create scoring and risk rules for low-risk approvals.
The example reflects the intended shift from conversational help toward bounded handoff. A person defines the goal and risk framework, then the model processes a complete workflow with review reserved for exceptions.
OpenAI says GPT-5.5 matches GPT-5.4’s per-token serving latency despite higher intelligence. That matters for interactive agents, although end-to-end speed also depends on tool calls and the number of reasoning steps.
Benchmarks indicate meaningful progress, particularly for terminals and long documents. They do not prove reliability inside every enterprise stack. Teams need sandboxed trials, error analysis and clear approval gates before granting broader authority.
GPT-5.5’s strongest proposition is fewer interventions per completed task. Its greatest challenge is showing that the reduction remains valuable after higher token prices, tool costs and governance are included.
If the model consistently finishes work that earlier systems left half-complete, the premium may be justified. The next evidence will come not from another score, but from production pipelines where a successful agent must be accurate, economical and trustworthy at the same time.
